SecureVector SDK for LangGraph

Bring the SecureVector local threat monitor's three controls — tool-call permissions, secret / data-leak detection, and threat detection — to every LangGraph tool call, with tamper-evident audit logging. One import.

pip install securevector-sdk-langgraph

📦 One install — batteries included. pip install securevector-sdk-langgraph also installs the local SecureVector app (securevector-ai-monitor): the adapter and the detection engine + tamper-evident audit chain arrive in a single pip install. The SDK is a thin interception layer — the app must be running locally (securevector-app --web) for it to do anything.

🌐 Pointing at your own cloud? Use the lightweight install. If you've deployed SecureVector to your own cloud, you don't need the bundled local app. Install only the adapter on the machine where your agents run, and point it at your deployment:

# lightweight — adapter only, no local app (your env already has langgraph)
pip install securevector-sdk-langgraph --no-deps

# point at your SecureVector endpoint — all you need for a private (in-VPC) endpoint
export SECUREVECTOR_SDK_APP_URL=https://<your-securevector-endpoint>

# OPTIONAL: only if your endpoint is publicly exposed and gated with an inbound token.
# A private endpoint in your own VPC needs no key. To gate a public one, use a free
# SecureVector cloud account API key or an SVET token — it gates access only; no agent
# data is sent to SecureVector.
export SECUREVECTOR_API_KEY=<SecureVector account key or SVET token>

The adapter then forwards every tool call to your remote deployment instead of a local app. The default pip install securevector-sdk-langgraph (no --no-deps) still bundles the app for local use.

Quick start

Enforcement (recommended) — the documented wrap_tool_call middleware, accepted by the langgraph-backed create_agent (note: langgraph.prebuilt.create_react_agent does not take a middleware argument — use create_agent):

from securevector_sdk_langgraph import secure_middleware
from langchain.agents import create_agent

agent = create_agent(
    model, tools,
    middleware=[secure_middleware(mode="enforce")],
)

A denied tool is short-circuited with a ToolMessage before it runs — no exceptions, no crashed graph.

Observe-only logging for any graph (passes through langchain-core's callback manager):

from securevector_sdk_langgraph import SecureVectorCallbackHandler

graph.invoke(state, config={"callbacks": [SecureVectorCallbackHandler()]})

Raw StateGraph with custom tool nodes (no middleware surface): gate the tool with LangGraph's documented interrupt() for human/programmatic approval:

from langgraph.types import interrupt

@tool
def run_query(sql: str):
    interrupt({"action": "run_query", "args": {"sql": sql}})  # pause for approval
    ...

Why these paths? LangGraph callbacks are observability-only — they cannot cleanly block a tool. The wrap_tool_call middleware (for create_agent) and interrupt() (for raw graphs) are the documented gates.

LLM cost tracking — add cost_tracking_middleware() and every model call's token usage (input / output / cached) posts to the app's Cost Tracking, where the pricing table turns it into real dollars (your agent runs on your own API keys) and per-agent budgets apply:

from securevector_sdk_langgraph import secure_middleware, cost_tracking_middleware

agent = create_agent(
    model, tools,
    middleware=[secure_middleware(mode="enforce"), cost_tracking_middleware()],
)

Records are attributed to agent_id="langgraph-agent" by default — name your agent with cost_tracking_middleware(agent_id="checkout-bot") or SECUREVECTOR_SDK_AGENT_ID. The callback handler captures the same usage via on_llm_end for raw graphs. Cost capture is best-effort: an unreachable app never breaks the agent.

What happens on every tool call

Before a tool node runs, the SDK:

1. (a) Permissions — resolves an allow/block verdict for the tool, using the app's own precedence: cloud-pushed synced policy → local override → essential registry → default-allow.
2. (b)+(c) Secret & threat scan — sends the serialized tool input through the app's /analyze pipeline.

After the tool returns, the result is scanned the same way to catch secrets / exfiltration in tool output. Every decision is written to the app's audit chain tagged runtime_kind="langgraph".

observe vs enforce

agent = create_agent(model, tools, middleware=[secure_middleware(mode="enforce")])

Enforce mode prints a one-time disclosure to stderr. (Enforcement requires the middleware or interrupt() path; the observe callback handler always logs only.)

Configuration

All optional, via env or install(...) kwargs:

Compliance

The tool-call-level, attributed, tamper-evident audit trail this produces is exactly the action-layer logging auditors ask for under EU AI Act Art. 12 / 15. This SDK produces the local evidence; the cloud governance surface turns it into an auditor-ready pack.

Trademarks

SecureVector is the product name of this SDK. LangGraph and LangChain are trademarks of LangChain, Inc. This is an independent, community SDK that integrates with LangGraph via its public callback API. It is not affiliated with, sponsored by, or endorsed by LangChain, Inc. The name uses "langgraph" only descriptively, to identify the framework this package works with (nominative fair use).

License

Apache-2.0. See LICENSE and NOTICE.